How AI uncovers new ways to tackle difficult diseases

During a video call, Alex Zhavoronkov held up a small, green, diamond-shaped pill. Developed by his company, the pill is intended to treat a rare, progressive lung disease for which the cause is unknown and no effective treatment exists. The new drug is not yet approved, but in small clinical trials, it has shown significant efficacy in treating idiopathic pulmonary fibrosis (IPF).

The drug is one of a new wave of medicines where artificial intelligence (AI) has played a crucial role in their discovery. “We can’t say we have the first AI-discovered and designed molecule approved,” said Dr. Zhavoronkov, co-founder and CEO of the US startup Insilico Medicine, “but we are probably the furthest along the road.” This marks the beginning of an AI drug race, with numerous companies harnessing the power of AI to accomplish tasks traditionally performed by medicinal chemists.

This includes smaller, specialist AI-driven biotech firms that have emerged over the last decade, as well as larger pharmaceutical companies that are either conducting their own research or partnering with smaller companies. Alphabet (Google’s parent company) is among the newer entrants, having launched AI drug discovery company Isomorphic Labs in the UK in late 2021. Its CEO, Demis Hassabis, shared this year’s Nobel Prize in Chemistry for his AI models with applications in AI drug design.

Chris Meier of the Boston Consulting Group (BCG) says that using AI for drug discovery could be a “huge game changer” for patients. It takes an average of 10 to 15 years and more than $2 billion (£1.6 billion) to bring a new drug to market. And the risks are high: about 90% of drugs that enter clinical trials fail. The hope is that using AI for drug discovery processes can reduce both time and costs, while also improving success rates.

Charlotte Deane, a professor of structural bioinformatics at Oxford University, says that a new era of AI-centered drug discovery is emerging. She has developed free AI tools to help pharmaceutical companies and other institutions improve drug discovery. “We’re at the beginning of understanding how good it can be,” she says. Experts say that this is unlikely to lead to a reduction in the number of pharmaceutical scientists, with the real savings coming from reducing the number of failures, but it will mean working in partnership with AI.

A recent analysis by BCG found that at least 75 “AI-discovered molecules” have entered clinical trials, with more expected. “They are now routinely going into clinical trials, which is a significant milestone,” Dr. Meier says. The next—and “bigger milestone”—will be when they start to show results. However, Professor Deane points out that there is no precise definition of an “AI-discovered” drug and that there is still a large degree of human involvement in all examples to date.

Dr. Meier explains that AI is primarily deployed in two steps of the drug discovery process. The first is identifying the therapeutic target at the molecular level, i.e., what a drug is designed to correct, such as a gene or protein that is altered in a way it shouldn’t be due to disease. Traditionally, scientists experimentally test potential targets in the lab based on their understanding of the disease, while AI can be trained to mine large databases to establish potential molecular biological links to disease and make suggestions.

The second, and more common, step is to design the drug to correct the target. This employs generative AI, which is also the basis of ChatGPT, to come up with molecules that could bind to the target and have an effect, replacing the expensive, manual process of chemists synthesizing hundreds of variants of the same molecule and trying to find the best one. Insilico Medicine, founded in 2014 and having secured more than $425 million in funding, uses AI in both steps and predicts the probability of clinical trial success, which is then fed back into its drug discovery work.

Currently, the company has six molecules in clinical trials, including the one for IPF, with next-stage trials being planned. In addition, four more molecules have been approved to enter trials, and nearly 30 more show promise. “All of these are ‘discovered from scratch using generative AI’,” says Dr. Zhavoronkov. “Our machines keep coming up with ideas until they come up with a perfect drug that meets all our criteria.”

The company's generative AI software designed a novel molecule to treat IPF after being given the goal of inhibiting a protein called TNIK. TNIK had never been used as a target for IPF before, but the company's other suite of AI software identified it as the most likely modulator of the disease. The possibilities suggested by the system were then synthesized and tested. Dr. Zhavoronkov notes that this discovery process was much faster and more streamlined than the industry standard.

It took 18 months and required the synthesis and testing of 79 molecules, compared with the usual expectation of around four years and at least 500 molecules synthesized. Insilico's other molecules have required even fewer, he says. Experts say that the lack of data from which AI can learn remains the biggest challenge facing the field. This applies to both target identification and molecule design, and can introduce bias.

US-based Recursion Pharmaceuticals says its approach mitigates the problem of limited data. By automating experiments, it generates vast amounts of data related to the entire molecular set-up that makes up the human body. It then trains AI tools to understand this data and find unexpected relationships. To help with this, the company installed what is claimed to be the fastest supercomputer owned and operated by any pharmaceutical company last year. It has had some success. One of its molecules, designed to treat lymphoma and solid tumors, is currently being tested in cancer patients in early clinical trials.

The molecule was developed after AI discovered a new way to target a gene thought to be important in driving these cancers, but which no one had previously been able to target on its own. Chris Gibson, co-founder and CEO of Recursion, says that the most important thing for the field is something that Recursion or anyone else has not yet demonstrated: that these AI-discovered molecules can get through clinical trials and, over time, provide a higher probability of success than traditional methods. When that happens, Dr. Gibson says, “the world will see very clearly that this is the way forward.”